Mr.

Inversion, 80’s – early 90’s: Albert Tarantola

• Basic properties of seismic inversion via least

squares and Newton’s method
• Practical algorithms for least-squares inversion
• Bayesian framework (“solution = a posteriori pdf”)
 Disaster!
• After a flurry of interest in the 80’s, industry
interest waned because…
• It didn’t work!
• Newton’s method converges to local min
poorly fitting data

Illustration based on Marmousi model…

 0%
surface position (km)
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0
0
0.5
depth (km)

1.0
1.5
2.0
2.5
3.0

10 20 30 40 50 60 70
Bulk Modulus (GPa)
 100 %
surface position (km)
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0
0
0.5
depth (km)

1.0
1.5
2.0
2.5
3.0

10 20 30 40 50 60 70
Bulk Modulus (GPa)
 95%
surface position (km)
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0
0
0.5
depth (km)

1.0
1.5
2.0
2.5
3.0

10 20 30 40 50 60 70
Bulk Modulus (GPa)
 90%
surface position (km)
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0
0
0.5
depth (km)

1.0
1.5
2.0
2.5
3.0

10 20 30 40 50 60 70
Bulk Modulus (GPa)
 80%
surface position (km)
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0
0
0.5
depth (km)

1.0
1.5
2.0
2.5
3.0

10 20 30 40 50 60 70
Bulk Modulus (GPa)
 70%
surface position (km)
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0
0
0.5
depth (km)

1.0
1.5
2.0
2.5
3.0

10 20 30 40 50 60 70
Bulk Modulus (GPa)
 100% - RMSE = 0%
offset (km) offset (km)
-2 -1 -2 -1
0 0
time (s)

time (s)
2 2

Shot record 121 – model 100% Data error

model 100% - model100%
 95% - RMSE = 184%
offset (km) offset (km)
-2 -1 -2 -1
0 0
time (s)

time (s)
2 2

Shot record 121 – model 95% Data error

model 95% - model100%
 90% - RMSE = 144%
offset (km) offset (km)
-2 -1 -2 -1
0 0
time (s)

time (s)
2 2

Shot record 121 – model 90% Data error

model 90% - model100%
 80% - RMSE = 179%
offset (km) offset (km)
-2 -1 -2 -1
0 0
time (s)

time (s)
2 2

Shot record 121 – model 80% Data error

model 80% - model 100%
 70% - RMSE = 216%
offset (km) offset (km)
-2 -1 -2 -1
0 0
time (s)

time (s)
2 2

Shot record 121 – model 70% Data error

model 70% - model 100%
 60% - RMSE = 273%
offset (km) offset (km)
-2 -1 -2 -1
0 0
time (s)

time (s)
2 2

Shot record 121 – model 60% Data error

model 60% - model 100%
• Kolb et al. 86: frequency continuation w low
starting freq increases chances of convergence
• Bunks et al. 95: success with Marmousi, very low
frequency data (0.25 Hz – compare typical 3-5 Hz)

Gerhard Pratt: many “algorithmic engineering”

contributions over the 90’s – exponential damping,
frequency decimation,
traveltime tomography
for initial models

Upshot: functional least-

squares inversion for
transmission data
• BP blind test at EAGE 04: Pratt’s result rekindles interest in
least-squares inversion by Newton
• now called “Full Waveform Inversion” (FWI)
• Every major firm has large team working on FWI
• Many successful field trials reported

Math has not changed since

Tarantola:

• Limited mostly to
transmission
• Requires very low
frequency data with good
s/n, or very good starting
model
(Brenders & Pratt, SEG 07)
 Origin of Extended Modeling
A dinner conversation in 1984:

Me: “Least squares inversion doesn’t work, whine,

whine”

Industry buddy: “We geophysicists find seismic models

thousands of times, every day, all over the world.
What’s wrong with you mathematicians?”

Me: “Ummm…”
 Extended Modeling and Inversion
Idea embedded in geophysical practice since
60’s, maybe before (Dobrin, p 234):

• Don’t need entire survey for inversion – can

estimate (eg.) one model per shot record – an
underdetermined problem!
 100% 90% 80%

williamsymes, Thu Jan 31 21:43

williamsymes, Thu Jan 31 21:43

Three inversions of shot 61 with different

starting models
 Extended Modeling and Inversion
• Select (somehow) an inversion for each shot

• Creates an extended model – depends on an

extra parameter (shot number or position),
fits data

• Special case – models same for all shots –

solution of original inverse problem!
williamsymes, Thu Jan 31 18:37

An extended inversion of Marmousi data

 Semblance
• There is only one earth: Amongst all extended
models fitting the data, choose one that isn’t
extended – all single-shot inversions same!

• Central issues: (i) how to navigate extended

models efficiently, (ii) how to measure semblance
= extent to which all models are same

• Like split-screen focusing

 100% 90% 80%

williamsymes, Thu Jan 31 22:05

williamsymes, Thu Jan 31 22:05

Slice of inverted extended model volumes as function of initial data

along shot axis for horizontal position 4.2 km – exhibits extent of
semblance violation
 Differential Semblance
• Measure degree of dependence on extra
param (shot) by differentiation
|F[c]-d|2 + α|Dsc|2
• Most studied variant: replace F[c] with F[v]r,
extend r only – then
minr [|F[v]r-d|2+α|Dsr|] = < d, P[v] d>
with P[v] = ΨDO dep smoothly on v
• A smoothly turning focusing knob!
BEFORE Seismic Autofocus by AFTER
Differential Semblance

Version developed in Peng

Shen’s PhD thesis: redundant
parameters via operator
coefficents in wave equation.

Applied to exploration
survey, southern Caribbean –
distortion of subsurface
structure due to gas
chimney. DS correctly locates
gas, focuses inversion to
reveal structure

[P. Shen & W. Symes, Geophysics 2008] – Thanks:

Shell
Review paper on FWI, velocity analysis,
semblance etc.: WWS, Inverse Problems, 2009

Many recent conference papers on extended

model inversion, including nonlinear version
(F[c] instead of F[v]r)

FWI without “low” frequencies appears feasible

– but theory needed!!!!
 Thanks to…
• students and collaborators
• Sponsors of The Rice Inversion Project
• Gunther, Laurent, Sean, Russ, Francois
• MSRI and NSF

And to all of you for listening!